Optical Switches Target Next-Generation AI and HPC Networks

As generative artificial intelligence and high-performance computing (HPC) advance at an unprecedented pace, data center networks are undergoing a profound transformation. The training of large language models, the scaling of inference clusters, and the extreme compute demands of scientific computing have made the network a critical bottleneck for overall system performance. Traditional electronic packet-switched architectures are increasingly struggling with bandwidth scaling, power consumption, latency, and reliability. Against this backdrop, optical circuit switches (OCS) and co-packaged optics (CPO) have rapidly moved from research laboratories into real-world deployments, emerging as key technologies for next-generation AI and HPC networks.

Core Challenges Facing AI and HPC Networks
Modern AI training work""s exhibit highly deterministic and periodic collective communication patterns. Whether using data parallelism, model parallelism, or pipeline parallelism, GPUs or TPUs must frequently exchange large volumes of synchronized data. These communication patterns impose stringent requirements on the network: high bandwidth, ultra-low latency, deterministic latency, and high reliability.

Traditional electronic switches require repeated optical-electrical-optical (O-E-O) conversions. Each port needs optical modules to perform electro-optic and opto-electric conversions, and power consumption rises sharply with higher data rates. Processing and queuing delays inherent to electronic switches are difficult to eliminate and can significantly degrade training efficiency in ultra-large-scale clusters. As cluster sizes grow from thousands to tens or even hundreds of thousands of GPUs, the number of switches and optical modules in conventional three-tier Fat-Tree topologies explodes. This drives up both capital expenditure (CapEx) and operational expenditure (OpEx), while creating severe thermal and power delivery challenges.

Moreover, electronic switches require hardware replacement with every generational speed upgrade, making cross-generational reuse difficult. AI and HPC networks demand extremely high operational stability— a brief network interruption can halt training jobs for hours or days, resulting in substantial economic losses. These practical constraints are accelerating the industry’s shift toward optical switching technologies.

Working Principles and Core Advantages of Optical Circuit Switches
An optical circuit switch performs signal path switching directly in the optical domain. It functions essentially as a programmable optical cross-connect. Rather than inspecting packet contents, it changes the physical path of optical signals through mechanical, micro-electromechanical systems (MEMS), silicon photonics, liquid crystal, or other mechanisms to establish all-optical connections between ports.

Compared with electronic packet switches, OCS offers several compelling advantages:
1. Extremely Low Power Consumption
With minimal need for O-E-O conversion or high-speed electronic processing, OCS port power can drop to tens of milliwatts or lower, while equivalent electronic switch ports typically consume more than 10 watts. Google has publicly reported approximately 40% power reduction and 30% cost savings in the relevant network segments after deploying OCS in its data centers.
2. Ultra-Low and Deterministic Latency
OCS latency is primarily determined by the propagation time of light within the device, typically in the tens of nanoseconds. Silicon photonics implementations can further reduce port-to-port latency to below 10 nanoseconds. In contrast, high-performance electronic switches still exhibit latencies of hundreds of nanoseconds, which worsen under congestion.
3. Rate and Protocol Transparency
OCS is completely transparent to signal rate, modulation format, and protocol. The same hardware can support 400G, 800G, and evolve smoothly to 1.6T or higher rates without replacement. This cross-generational reusability gives OCS long-term infrastructure-like value.
4. High Reliability and Fast Failure Recovery
OCS systems contain far fewer active electronic components than traditional switches, resulting in fewer points of failure. Combined with redundant ports and software control, path switching can be completed in milliseconds or less, significantly improving network availability. Google has indicated that its OCS networks have achieved roughly 50 times better long-term uptime.
5. Reconfigurable Network Topology
OCS enables software-driven dynamic adjustment of network topology to match the communication patterns of different training jobs. This capability is particularly valuable for AI work""s, improving GPU utilization and reducing training time.

Key Application Scenarios in AI and HPC Networks
Current OCS deployments focus primarily on two layers:
· Scale-Out (Horizontal Cluster Expansion)
This is currently the most mature application for OCS. Google pioneered large-scale deployment of its own OCS designs in TPU supercomputers, using them to replace traditional spine-layer switches. Dynamic topology reconfiguration allows the system to more efficiently support training jobs of varying scales while reducing power and cost. Market analysis from Cignal AI indicates that scale-out remains the primary volume driver for OCS shipments in the near term.

· Scale-Up (Vertical System Expansion)
As GPU interconnects expand from within a single rack to multi-rack or larger domains (for example, NVIDIA’s NVL72 and NVL576 architectures), the demand for low-latency, high-bandwidth memory-class interconnects becomes critical. Traditional copper links are approaching their limits in both distance and speed. Silicon photonics OCS, with its ultra-low latency and scalable port counts, is emerging as a strong candidate for scale-up networks. Companies such as Salience Labs have introduced silicon photonics OCS with port-to-port latencies on the order of 10 nanoseconds—significantly better than conventional Ethernet switches—potentially enabling higher training throughput while remaining within a single system architecture.

OCS is also being applied to improve network resilience (rapid failover around failed links), support reconfigurable topologies, and partially replace electronic switches in hybrid electro-optical architectures to reduce overall power consumption and module counts.

Industry Progress and Key Players
Optical switching technology has progressed from concept validation to scaled deployment and commercialization.
· Google: As the pioneer of OCS in data center applications, the company has invested more than $1 billion and publicly demonstrated significant gains in power efficiency and reliability.
· NVIDIA: Advancing CPO technology through its Quantum-X Photonics InfiniBand and Spectrum-X Photonics Ethernet switches, with commercial availability targeted for 2026. CPO integrates optical engines directly onto the switch ASIC, substantially reducing power, latency, and component count while improving system resilience.
· Huawei: Launched the OptiXtrans DC808 all-optical switch, which uses MEMS technology to deliver high-density, low-power all-optical cross-connects that support smooth evolution to higher rates and has received recognition at events such as Interop.
· Molex: Introduced a high-radix OCS platform leveraging long-standing MEMS expertise to provide reconfigurable optical connectivity for AI clusters.
· Startups: Salience Labs focuses on silicon photonics OCS with emphasis on ultra-low latency for scale-up applications; Finchetto is exploring true optical packet switching to eliminate electronic control bottlenecks; Lumotive and others are pursuing higher port densities using optical metamaterials and related approaches.

Technology approaches remain diverse: MEMS continues to dominate high-port-count solutions, silicon photonics offers advantages in latency and integration, while liquid crystal, piezoelectric, and other technologies find niches in specific use cases. Market forecasts from Cignal AI project the OCS market will exceed $8 billion by 2030, with scale-up (GPU) demand expected to become a major growth driver.

Challenges and the Trend Toward Hybrid Architectures
· Despite the promising outlook, widespread adoption of optical switches still faces several challenges:
· Reconfiguration time: Traditional MEMS solutions typically operate in the millisecond range, which remains suboptimal for scenarios requiring frequent switching. Newer silicon photonics approaches are driving reconfiguration times lower.
· Insertion loss and optical power budget: Controlling loss in high-port-count OCS remains an engineering challenge.
· Software and control plane maturity: Robust management systems are needed for topology optimization, fault detection, and work"" matching.
· Coexistence with existing electronic networks: Fully optical networks are unlikely to completely displace electronic switches in the short term; hybrid electro-optical architectures will serve as the primary transitional model.

Industry consensus holds that OCS will first achieve large-scale deployment in the scale-out spine layer and selected scale-up scenarios before penetrating deeper into the network. CPO technology, working from inside the switch, will reduce reliance on traditional pluggable optics and jointly drive networks toward higher efficiency.

Outlook: Optical Networks as the Foundation of AI Infrastructure
Under the dual pressures of rising energy costs and carbon-neutrality goals, data centers have elevated energy efficiency to a strategic priority. Optical switches directly address this need by eliminating unnecessary O-E-O conversions, reducing component counts, and enabling topology optimization. At the same time, their rate transparency provides long-term reusability that helps lower total cost of ownership.

In the coming years, the majority of interconnects in AI data centers are expected to accelerate their transition to optical solutions. The combination of OCS and CPO will not only resolve current power and scaling bottlenecks but also provide a robust network foundation for ever-larger “AI factories” and HPC systems. Optical switches have evolved from an optional technology into a critical enabler for next-generation AI and HPC networks, redefining the boundaries of energy efficiency, latency, and scalability in computing infrastructure.